1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2015 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * on the rights to use, copy, modify, merge, publish, distribute, sub
10 * license, and/or sell copies of the Software, and to permit persons to whom
11 * the Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
21 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
22 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 #include "si_compute.h"
27 #include "si_pipe.h"
28 #include "util/format/u_format.h"
29 #include "util/u_log.h"
30 #include "util/u_surface.h"
31
32 enum
33 {
34 SI_COPY =
35 SI_SAVE_FRAMEBUFFER | SI_SAVE_TEXTURES | SI_SAVE_FRAGMENT_STATE | SI_DISABLE_RENDER_COND,
36
37 SI_BLIT = SI_SAVE_FRAMEBUFFER | SI_SAVE_TEXTURES | SI_SAVE_FRAGMENT_STATE,
38
39 SI_DECOMPRESS = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE | SI_DISABLE_RENDER_COND,
40
41 SI_COLOR_RESOLVE = SI_SAVE_FRAMEBUFFER | SI_SAVE_FRAGMENT_STATE
42 };
43
si_blitter_begin(struct si_context * sctx,enum si_blitter_op op)44 void si_blitter_begin(struct si_context *sctx, enum si_blitter_op op)
45 {
46 util_blitter_save_vertex_shader(sctx->blitter, sctx->shader.vs.cso);
47 util_blitter_save_tessctrl_shader(sctx->blitter, sctx->shader.tcs.cso);
48 util_blitter_save_tesseval_shader(sctx->blitter, sctx->shader.tes.cso);
49 util_blitter_save_geometry_shader(sctx->blitter, sctx->shader.gs.cso);
50 util_blitter_save_so_targets(sctx->blitter, sctx->streamout.num_targets,
51 (struct pipe_stream_output_target **)sctx->streamout.targets);
52 util_blitter_save_rasterizer(sctx->blitter, sctx->queued.named.rasterizer);
53
54 if (op & SI_SAVE_FRAGMENT_STATE) {
55 util_blitter_save_blend(sctx->blitter, sctx->queued.named.blend);
56 util_blitter_save_depth_stencil_alpha(sctx->blitter, sctx->queued.named.dsa);
57 util_blitter_save_stencil_ref(sctx->blitter, &sctx->stencil_ref.state);
58 util_blitter_save_fragment_shader(sctx->blitter, sctx->shader.ps.cso);
59 util_blitter_save_sample_mask(sctx->blitter, sctx->sample_mask);
60 util_blitter_save_scissor(sctx->blitter, &sctx->scissors[0]);
61 util_blitter_save_window_rectangles(sctx->blitter, sctx->window_rectangles_include,
62 sctx->num_window_rectangles, sctx->window_rectangles);
63 }
64
65 if (op & SI_SAVE_FRAMEBUFFER)
66 util_blitter_save_framebuffer(sctx->blitter, &sctx->framebuffer.state);
67
68 if (op & SI_SAVE_TEXTURES) {
69 util_blitter_save_fragment_sampler_states(
70 sctx->blitter, 2, (void **)sctx->samplers[PIPE_SHADER_FRAGMENT].sampler_states);
71
72 util_blitter_save_fragment_sampler_views(sctx->blitter, 2,
73 sctx->samplers[PIPE_SHADER_FRAGMENT].views);
74 }
75
76 if (op & SI_DISABLE_RENDER_COND)
77 sctx->render_cond_enabled = false;
78
79 if (sctx->screen->dpbb_allowed) {
80 sctx->dpbb_force_off = true;
81 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
82 }
83
84 sctx->blitter_running = true;
85 }
86
si_blitter_end(struct si_context * sctx)87 void si_blitter_end(struct si_context *sctx)
88 {
89 sctx->blitter_running = false;
90
91 if (sctx->screen->dpbb_allowed) {
92 sctx->dpbb_force_off = false;
93 si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
94 }
95
96 sctx->render_cond_enabled = sctx->render_cond;
97
98 /* Restore shader pointers because the VS blit shader changed all
99 * non-global VS user SGPRs. */
100 sctx->shader_pointers_dirty |= SI_DESCS_SHADER_MASK(VERTEX);
101
102 unsigned num_vbos_in_user_sgprs = si_num_vbos_in_user_sgprs(sctx->screen);
103 sctx->vertex_buffer_pointer_dirty = sctx->vb_descriptors_buffer != NULL &&
104 sctx->num_vertex_elements >
105 num_vbos_in_user_sgprs;
106 sctx->vertex_buffer_user_sgprs_dirty = sctx->num_vertex_elements > 0 &&
107 num_vbos_in_user_sgprs;
108 si_mark_atom_dirty(sctx, &sctx->atoms.s.shader_pointers);
109 }
110
u_max_sample(struct pipe_resource * r)111 static unsigned u_max_sample(struct pipe_resource *r)
112 {
113 return r->nr_samples ? r->nr_samples - 1 : 0;
114 }
115
si_blit_dbcb_copy(struct si_context * sctx,struct si_texture * src,struct si_texture * dst,unsigned planes,unsigned level_mask,unsigned first_layer,unsigned last_layer,unsigned first_sample,unsigned last_sample)116 static unsigned si_blit_dbcb_copy(struct si_context *sctx, struct si_texture *src,
117 struct si_texture *dst, unsigned planes, unsigned level_mask,
118 unsigned first_layer, unsigned last_layer, unsigned first_sample,
119 unsigned last_sample)
120 {
121 struct pipe_surface surf_tmpl = {{0}};
122 unsigned layer, sample, checked_last_layer, max_layer;
123 unsigned fully_copied_levels = 0;
124
125 if (planes & PIPE_MASK_Z)
126 sctx->dbcb_depth_copy_enabled = true;
127 if (planes & PIPE_MASK_S)
128 sctx->dbcb_stencil_copy_enabled = true;
129 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
130
131 assert(sctx->dbcb_depth_copy_enabled || sctx->dbcb_stencil_copy_enabled);
132
133 sctx->decompression_enabled = true;
134
135 while (level_mask) {
136 unsigned level = u_bit_scan(&level_mask);
137
138 /* The smaller the mipmap level, the less layers there are
139 * as far as 3D textures are concerned. */
140 max_layer = util_max_layer(&src->buffer.b.b, level);
141 checked_last_layer = MIN2(last_layer, max_layer);
142
143 surf_tmpl.u.tex.level = level;
144
145 for (layer = first_layer; layer <= checked_last_layer; layer++) {
146 struct pipe_surface *zsurf, *cbsurf;
147
148 surf_tmpl.format = src->buffer.b.b.format;
149 surf_tmpl.u.tex.first_layer = layer;
150 surf_tmpl.u.tex.last_layer = layer;
151
152 zsurf = sctx->b.create_surface(&sctx->b, &src->buffer.b.b, &surf_tmpl);
153
154 surf_tmpl.format = dst->buffer.b.b.format;
155 cbsurf = sctx->b.create_surface(&sctx->b, &dst->buffer.b.b, &surf_tmpl);
156
157 for (sample = first_sample; sample <= last_sample; sample++) {
158 if (sample != sctx->dbcb_copy_sample) {
159 sctx->dbcb_copy_sample = sample;
160 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
161 }
162
163 si_blitter_begin(sctx, SI_DECOMPRESS);
164 util_blitter_custom_depth_stencil(sctx->blitter, zsurf, cbsurf, 1 << sample,
165 sctx->custom_dsa_flush, 1.0f);
166 si_blitter_end(sctx);
167 }
168
169 pipe_surface_reference(&zsurf, NULL);
170 pipe_surface_reference(&cbsurf, NULL);
171 }
172
173 if (first_layer == 0 && last_layer >= max_layer && first_sample == 0 &&
174 last_sample >= u_max_sample(&src->buffer.b.b))
175 fully_copied_levels |= 1u << level;
176 }
177
178 sctx->decompression_enabled = false;
179 sctx->dbcb_depth_copy_enabled = false;
180 sctx->dbcb_stencil_copy_enabled = false;
181 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
182
183 return fully_copied_levels;
184 }
185
186 /* Helper function for si_blit_decompress_zs_in_place.
187 */
si_blit_decompress_zs_planes_in_place(struct si_context * sctx,struct si_texture * texture,unsigned planes,unsigned level_mask,unsigned first_layer,unsigned last_layer)188 static void si_blit_decompress_zs_planes_in_place(struct si_context *sctx,
189 struct si_texture *texture, unsigned planes,
190 unsigned level_mask, unsigned first_layer,
191 unsigned last_layer)
192 {
193 struct pipe_surface *zsurf, surf_tmpl = {{0}};
194 unsigned layer, max_layer, checked_last_layer;
195 unsigned fully_decompressed_mask = 0;
196
197 if (!level_mask)
198 return;
199
200 if (planes & PIPE_MASK_S)
201 sctx->db_flush_stencil_inplace = true;
202 if (planes & PIPE_MASK_Z)
203 sctx->db_flush_depth_inplace = true;
204 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
205
206 surf_tmpl.format = texture->buffer.b.b.format;
207
208 sctx->decompression_enabled = true;
209
210 while (level_mask) {
211 unsigned level = u_bit_scan(&level_mask);
212
213 surf_tmpl.u.tex.level = level;
214
215 /* The smaller the mipmap level, the less layers there are
216 * as far as 3D textures are concerned. */
217 max_layer = util_max_layer(&texture->buffer.b.b, level);
218 checked_last_layer = MIN2(last_layer, max_layer);
219
220 for (layer = first_layer; layer <= checked_last_layer; layer++) {
221 surf_tmpl.u.tex.first_layer = layer;
222 surf_tmpl.u.tex.last_layer = layer;
223
224 zsurf = sctx->b.create_surface(&sctx->b, &texture->buffer.b.b, &surf_tmpl);
225
226 si_blitter_begin(sctx, SI_DECOMPRESS);
227 util_blitter_custom_depth_stencil(sctx->blitter, zsurf, NULL, ~0, sctx->custom_dsa_flush,
228 1.0f);
229 si_blitter_end(sctx);
230
231 pipe_surface_reference(&zsurf, NULL);
232 }
233
234 /* The texture will always be dirty if some layers aren't flushed.
235 * I don't think this case occurs often though. */
236 if (first_layer == 0 && last_layer >= max_layer) {
237 fully_decompressed_mask |= 1u << level;
238 }
239 }
240
241 if (planes & PIPE_MASK_Z)
242 texture->dirty_level_mask &= ~fully_decompressed_mask;
243 if (planes & PIPE_MASK_S)
244 texture->stencil_dirty_level_mask &= ~fully_decompressed_mask;
245
246 sctx->decompression_enabled = false;
247 sctx->db_flush_depth_inplace = false;
248 sctx->db_flush_stencil_inplace = false;
249 si_mark_atom_dirty(sctx, &sctx->atoms.s.db_render_state);
250 }
251
252 /* Helper function of si_flush_depth_texture: decompress the given levels
253 * of Z and/or S planes in place.
254 */
si_blit_decompress_zs_in_place(struct si_context * sctx,struct si_texture * texture,unsigned levels_z,unsigned levels_s,unsigned first_layer,unsigned last_layer)255 static void si_blit_decompress_zs_in_place(struct si_context *sctx, struct si_texture *texture,
256 unsigned levels_z, unsigned levels_s,
257 unsigned first_layer, unsigned last_layer)
258 {
259 unsigned both = levels_z & levels_s;
260
261 /* First, do combined Z & S decompresses for levels that need it. */
262 if (both) {
263 si_blit_decompress_zs_planes_in_place(sctx, texture, PIPE_MASK_Z | PIPE_MASK_S, both,
264 first_layer, last_layer);
265 levels_z &= ~both;
266 levels_s &= ~both;
267 }
268
269 /* Now do separate Z and S decompresses. */
270 if (levels_z) {
271 si_blit_decompress_zs_planes_in_place(sctx, texture, PIPE_MASK_Z, levels_z, first_layer,
272 last_layer);
273 }
274
275 if (levels_s) {
276 si_blit_decompress_zs_planes_in_place(sctx, texture, PIPE_MASK_S, levels_s, first_layer,
277 last_layer);
278 }
279 }
280
si_decompress_depth(struct si_context * sctx,struct si_texture * tex,unsigned required_planes,unsigned first_level,unsigned last_level,unsigned first_layer,unsigned last_layer)281 static void si_decompress_depth(struct si_context *sctx, struct si_texture *tex,
282 unsigned required_planes, unsigned first_level, unsigned last_level,
283 unsigned first_layer, unsigned last_layer)
284 {
285 unsigned inplace_planes = 0;
286 unsigned copy_planes = 0;
287 unsigned level_mask = u_bit_consecutive(first_level, last_level - first_level + 1);
288 unsigned levels_z = 0;
289 unsigned levels_s = 0;
290
291 if (required_planes & PIPE_MASK_Z) {
292 levels_z = level_mask & tex->dirty_level_mask;
293
294 if (levels_z) {
295 if (si_can_sample_zs(tex, false))
296 inplace_planes |= PIPE_MASK_Z;
297 else
298 copy_planes |= PIPE_MASK_Z;
299 }
300 }
301 if (required_planes & PIPE_MASK_S) {
302 levels_s = level_mask & tex->stencil_dirty_level_mask;
303
304 if (levels_s) {
305 if (si_can_sample_zs(tex, true))
306 inplace_planes |= PIPE_MASK_S;
307 else
308 copy_planes |= PIPE_MASK_S;
309 }
310 }
311
312 if (unlikely(sctx->log))
313 u_log_printf(sctx->log,
314 "\n------------------------------------------------\n"
315 "Decompress Depth (levels %u - %u, levels Z: 0x%x S: 0x%x)\n\n",
316 first_level, last_level, levels_z, levels_s);
317
318 /* We may have to allocate the flushed texture here when called from
319 * si_decompress_subresource.
320 */
321 if (copy_planes &&
322 (tex->flushed_depth_texture || si_init_flushed_depth_texture(&sctx->b, &tex->buffer.b.b))) {
323 struct si_texture *dst = tex->flushed_depth_texture;
324 unsigned fully_copied_levels;
325 unsigned levels = 0;
326
327 assert(tex->flushed_depth_texture);
328
329 if (util_format_is_depth_and_stencil(dst->buffer.b.b.format))
330 copy_planes = PIPE_MASK_Z | PIPE_MASK_S;
331
332 if (copy_planes & PIPE_MASK_Z) {
333 levels |= levels_z;
334 levels_z = 0;
335 }
336 if (copy_planes & PIPE_MASK_S) {
337 levels |= levels_s;
338 levels_s = 0;
339 }
340
341 fully_copied_levels = si_blit_dbcb_copy(sctx, tex, dst, copy_planes, levels, first_layer,
342 last_layer, 0, u_max_sample(&tex->buffer.b.b));
343
344 if (copy_planes & PIPE_MASK_Z)
345 tex->dirty_level_mask &= ~fully_copied_levels;
346 if (copy_planes & PIPE_MASK_S)
347 tex->stencil_dirty_level_mask &= ~fully_copied_levels;
348 }
349
350 if (inplace_planes) {
351 bool has_htile = si_htile_enabled(tex, first_level, inplace_planes);
352 bool tc_compat_htile = vi_tc_compat_htile_enabled(tex, first_level, inplace_planes);
353
354 /* Don't decompress if there is no HTILE or when HTILE is
355 * TC-compatible. */
356 if (has_htile && !tc_compat_htile) {
357 si_blit_decompress_zs_in_place(sctx, tex, levels_z, levels_s, first_layer, last_layer);
358 } else {
359 /* This is only a cache flush.
360 *
361 * Only clear the mask that we are flushing, because
362 * si_make_DB_shader_coherent() treats different levels
363 * and depth and stencil differently.
364 */
365 if (inplace_planes & PIPE_MASK_Z)
366 tex->dirty_level_mask &= ~levels_z;
367 if (inplace_planes & PIPE_MASK_S)
368 tex->stencil_dirty_level_mask &= ~levels_s;
369 }
370
371 /* We just had to completely decompress Z/S for texturing. Enable
372 * TC-compatible HTILE on the next clear, so that the decompression
373 * doesn't have to be done for this texture ever again.
374 *
375 * TC-compatible HTILE might slightly reduce Z/S performance, but
376 * the decompression is much worse.
377 */
378 if (has_htile && !tc_compat_htile &&
379 /* We can only transition the whole buffer in one clear, so no mipmapping: */
380 tex->buffer.b.b.last_level == 0 &&
381 tex->surface.flags & RADEON_SURF_TC_COMPATIBLE_HTILE &&
382 (inplace_planes & PIPE_MASK_Z || !tex->htile_stencil_disabled))
383 tex->enable_tc_compatible_htile_next_clear = true;
384
385 /* Only in-place decompression needs to flush DB caches, or
386 * when we don't decompress but TC-compatible planes are dirty.
387 */
388 si_make_DB_shader_coherent(sctx, tex->buffer.b.b.nr_samples, inplace_planes & PIPE_MASK_S,
389 tc_compat_htile);
390 }
391 /* set_framebuffer_state takes care of coherency for single-sample.
392 * The DB->CB copy uses CB for the final writes.
393 */
394 if (copy_planes && tex->buffer.b.b.nr_samples > 1)
395 si_make_CB_shader_coherent(sctx, tex->buffer.b.b.nr_samples, false, true /* no DCC */);
396 }
397
si_decompress_sampler_depth_textures(struct si_context * sctx,struct si_samplers * textures)398 static bool si_decompress_sampler_depth_textures(struct si_context *sctx,
399 struct si_samplers *textures)
400 {
401 unsigned i;
402 unsigned mask = textures->needs_depth_decompress_mask;
403 bool need_flush = false;
404
405 while (mask) {
406 struct pipe_sampler_view *view;
407 struct si_sampler_view *sview;
408 struct si_texture *tex;
409
410 i = u_bit_scan(&mask);
411
412 view = textures->views[i];
413 assert(view);
414 sview = (struct si_sampler_view *)view;
415
416 tex = (struct si_texture *)view->texture;
417 assert(tex->db_compatible);
418
419 si_decompress_depth(sctx, tex, sview->is_stencil_sampler ? PIPE_MASK_S : PIPE_MASK_Z,
420 view->u.tex.first_level, view->u.tex.last_level, 0,
421 util_max_layer(&tex->buffer.b.b, view->u.tex.first_level));
422
423 if (tex->need_flush_after_depth_decompression) {
424 need_flush = true;
425 tex->need_flush_after_depth_decompression = false;
426 }
427 }
428
429 return need_flush;
430 }
431
si_blit_decompress_color(struct si_context * sctx,struct si_texture * tex,unsigned first_level,unsigned last_level,unsigned first_layer,unsigned last_layer,bool need_dcc_decompress,bool need_fmask_expand)432 static void si_blit_decompress_color(struct si_context *sctx, struct si_texture *tex,
433 unsigned first_level, unsigned last_level,
434 unsigned first_layer, unsigned last_layer,
435 bool need_dcc_decompress, bool need_fmask_expand)
436 {
437 void *custom_blend;
438 unsigned layer, checked_last_layer, max_layer;
439 unsigned level_mask = u_bit_consecutive(first_level, last_level - first_level + 1);
440
441 if (!need_dcc_decompress)
442 level_mask &= tex->dirty_level_mask;
443 if (!level_mask)
444 goto expand_fmask;
445
446 if (unlikely(sctx->log))
447 u_log_printf(sctx->log,
448 "\n------------------------------------------------\n"
449 "Decompress Color (levels %u - %u, mask 0x%x)\n\n",
450 first_level, last_level, level_mask);
451
452 if (need_dcc_decompress) {
453 assert(sctx->chip_class == GFX8 || tex->buffer.b.b.nr_storage_samples >= 2);
454 custom_blend = sctx->custom_blend_dcc_decompress;
455
456 assert(vi_dcc_enabled(tex, first_level));
457
458 /* disable levels without DCC */
459 for (int i = first_level; i <= last_level; i++) {
460 if (!vi_dcc_enabled(tex, i))
461 level_mask &= ~(1 << i);
462 }
463 } else if (tex->surface.fmask_size) {
464 custom_blend = sctx->custom_blend_fmask_decompress;
465 } else {
466 custom_blend = sctx->custom_blend_eliminate_fastclear;
467 }
468
469 sctx->decompression_enabled = true;
470
471 while (level_mask) {
472 unsigned level = u_bit_scan(&level_mask);
473
474 /* The smaller the mipmap level, the less layers there are
475 * as far as 3D textures are concerned. */
476 max_layer = util_max_layer(&tex->buffer.b.b, level);
477 checked_last_layer = MIN2(last_layer, max_layer);
478
479 for (layer = first_layer; layer <= checked_last_layer; layer++) {
480 struct pipe_surface *cbsurf, surf_tmpl;
481
482 surf_tmpl.format = tex->buffer.b.b.format;
483 surf_tmpl.u.tex.level = level;
484 surf_tmpl.u.tex.first_layer = layer;
485 surf_tmpl.u.tex.last_layer = layer;
486 cbsurf = sctx->b.create_surface(&sctx->b, &tex->buffer.b.b, &surf_tmpl);
487
488 /* Required before and after FMASK and DCC_DECOMPRESS. */
489 if (custom_blend == sctx->custom_blend_fmask_decompress ||
490 custom_blend == sctx->custom_blend_dcc_decompress)
491 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
492
493 si_blitter_begin(sctx, SI_DECOMPRESS);
494 util_blitter_custom_color(sctx->blitter, cbsurf, custom_blend);
495 si_blitter_end(sctx);
496
497 if (custom_blend == sctx->custom_blend_fmask_decompress ||
498 custom_blend == sctx->custom_blend_dcc_decompress)
499 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
500
501 /* When running FMASK decompresion with DCC, we need to run the "eliminate fast clear" pass
502 * separately because FMASK decompression doesn't eliminate DCC fast clear. This makes
503 * render->texture transitions more expensive. It can be disabled by
504 * allow_dcc_msaa_clear_to_reg_for_bpp.
505 *
506 * TODO: When we get here, change the compression to TC-compatible on the next clear
507 * to disable both the FMASK decompression and fast clear elimination passes.
508 */
509 if (sctx->screen->allow_dcc_msaa_clear_to_reg_for_bpp[util_logbase2(tex->surface.bpe)] &&
510 custom_blend == sctx->custom_blend_fmask_decompress &&
511 vi_dcc_enabled(tex, level)) {
512 si_blitter_begin(sctx, SI_DECOMPRESS);
513 util_blitter_custom_color(sctx->blitter, cbsurf, sctx->custom_blend_eliminate_fastclear);
514 si_blitter_end(sctx);
515 }
516
517 pipe_surface_reference(&cbsurf, NULL);
518 }
519
520 /* The texture will always be dirty if some layers aren't flushed.
521 * I don't think this case occurs often though. */
522 if (first_layer == 0 && last_layer >= max_layer) {
523 tex->dirty_level_mask &= ~(1 << level);
524 }
525 }
526
527 sctx->decompression_enabled = false;
528 si_make_CB_shader_coherent(sctx, tex->buffer.b.b.nr_samples, vi_dcc_enabled(tex, first_level),
529 tex->surface.u.gfx9.color.dcc.pipe_aligned);
530
531 expand_fmask:
532 if (need_fmask_expand && tex->surface.fmask_offset && !tex->fmask_is_identity) {
533 si_compute_expand_fmask(&sctx->b, &tex->buffer.b.b);
534 tex->fmask_is_identity = true;
535 }
536 }
537
si_decompress_color_texture(struct si_context * sctx,struct si_texture * tex,unsigned first_level,unsigned last_level,bool need_fmask_expand)538 static void si_decompress_color_texture(struct si_context *sctx, struct si_texture *tex,
539 unsigned first_level, unsigned last_level,
540 bool need_fmask_expand)
541 {
542 /* CMASK or DCC can be discarded and we can still end up here. */
543 if (!tex->cmask_buffer && !tex->surface.fmask_size &&
544 !vi_dcc_enabled(tex, first_level))
545 return;
546
547 si_blit_decompress_color(sctx, tex, first_level, last_level, 0,
548 util_max_layer(&tex->buffer.b.b, first_level), false,
549 need_fmask_expand);
550 }
551
si_decompress_sampler_color_textures(struct si_context * sctx,struct si_samplers * textures)552 static void si_decompress_sampler_color_textures(struct si_context *sctx,
553 struct si_samplers *textures)
554 {
555 unsigned i;
556 unsigned mask = textures->needs_color_decompress_mask;
557
558 while (mask) {
559 struct pipe_sampler_view *view;
560 struct si_texture *tex;
561
562 i = u_bit_scan(&mask);
563
564 view = textures->views[i];
565 assert(view);
566
567 tex = (struct si_texture *)view->texture;
568
569 si_decompress_color_texture(sctx, tex, view->u.tex.first_level, view->u.tex.last_level,
570 false);
571 }
572 }
573
si_decompress_image_color_textures(struct si_context * sctx,struct si_images * images)574 static void si_decompress_image_color_textures(struct si_context *sctx, struct si_images *images)
575 {
576 unsigned i;
577 unsigned mask = images->needs_color_decompress_mask;
578
579 while (mask) {
580 const struct pipe_image_view *view;
581 struct si_texture *tex;
582
583 i = u_bit_scan(&mask);
584
585 view = &images->views[i];
586 assert(view->resource->target != PIPE_BUFFER);
587
588 tex = (struct si_texture *)view->resource;
589
590 si_decompress_color_texture(sctx, tex, view->u.tex.level, view->u.tex.level,
591 view->access & PIPE_IMAGE_ACCESS_WRITE);
592 }
593 }
594
si_check_render_feedback_texture(struct si_context * sctx,struct si_texture * tex,unsigned first_level,unsigned last_level,unsigned first_layer,unsigned last_layer)595 static void si_check_render_feedback_texture(struct si_context *sctx, struct si_texture *tex,
596 unsigned first_level, unsigned last_level,
597 unsigned first_layer, unsigned last_layer)
598 {
599 bool render_feedback = false;
600
601 if (!vi_dcc_enabled(tex, first_level))
602 return;
603
604 for (unsigned j = 0; j < sctx->framebuffer.state.nr_cbufs; ++j) {
605 struct si_surface *surf;
606
607 if (!sctx->framebuffer.state.cbufs[j])
608 continue;
609
610 surf = (struct si_surface *)sctx->framebuffer.state.cbufs[j];
611
612 if (tex == (struct si_texture *)surf->base.texture && surf->base.u.tex.level >= first_level &&
613 surf->base.u.tex.level <= last_level && surf->base.u.tex.first_layer <= last_layer &&
614 surf->base.u.tex.last_layer >= first_layer) {
615 render_feedback = true;
616 break;
617 }
618 }
619
620 if (render_feedback)
621 si_texture_disable_dcc(sctx, tex);
622 }
623
si_check_render_feedback_textures(struct si_context * sctx,struct si_samplers * textures,uint32_t in_use_mask)624 static void si_check_render_feedback_textures(struct si_context *sctx, struct si_samplers *textures,
625 uint32_t in_use_mask)
626 {
627 uint32_t mask = textures->enabled_mask & in_use_mask;
628
629 while (mask) {
630 const struct pipe_sampler_view *view;
631 struct si_texture *tex;
632
633 unsigned i = u_bit_scan(&mask);
634
635 view = textures->views[i];
636 if (view->texture->target == PIPE_BUFFER)
637 continue;
638
639 tex = (struct si_texture *)view->texture;
640
641 si_check_render_feedback_texture(sctx, tex, view->u.tex.first_level, view->u.tex.last_level,
642 view->u.tex.first_layer, view->u.tex.last_layer);
643 }
644 }
645
si_check_render_feedback_images(struct si_context * sctx,struct si_images * images,uint32_t in_use_mask)646 static void si_check_render_feedback_images(struct si_context *sctx, struct si_images *images,
647 uint32_t in_use_mask)
648 {
649 uint32_t mask = images->enabled_mask & in_use_mask;
650
651 while (mask) {
652 const struct pipe_image_view *view;
653 struct si_texture *tex;
654
655 unsigned i = u_bit_scan(&mask);
656
657 view = &images->views[i];
658 if (view->resource->target == PIPE_BUFFER)
659 continue;
660
661 tex = (struct si_texture *)view->resource;
662
663 si_check_render_feedback_texture(sctx, tex, view->u.tex.level, view->u.tex.level,
664 view->u.tex.first_layer, view->u.tex.last_layer);
665 }
666 }
667
si_check_render_feedback_resident_textures(struct si_context * sctx)668 static void si_check_render_feedback_resident_textures(struct si_context *sctx)
669 {
670 util_dynarray_foreach (&sctx->resident_tex_handles, struct si_texture_handle *, tex_handle) {
671 struct pipe_sampler_view *view;
672 struct si_texture *tex;
673
674 view = (*tex_handle)->view;
675 if (view->texture->target == PIPE_BUFFER)
676 continue;
677
678 tex = (struct si_texture *)view->texture;
679
680 si_check_render_feedback_texture(sctx, tex, view->u.tex.first_level, view->u.tex.last_level,
681 view->u.tex.first_layer, view->u.tex.last_layer);
682 }
683 }
684
si_check_render_feedback_resident_images(struct si_context * sctx)685 static void si_check_render_feedback_resident_images(struct si_context *sctx)
686 {
687 util_dynarray_foreach (&sctx->resident_img_handles, struct si_image_handle *, img_handle) {
688 struct pipe_image_view *view;
689 struct si_texture *tex;
690
691 view = &(*img_handle)->view;
692 if (view->resource->target == PIPE_BUFFER)
693 continue;
694
695 tex = (struct si_texture *)view->resource;
696
697 si_check_render_feedback_texture(sctx, tex, view->u.tex.level, view->u.tex.level,
698 view->u.tex.first_layer, view->u.tex.last_layer);
699 }
700 }
701
si_check_render_feedback(struct si_context * sctx)702 static void si_check_render_feedback(struct si_context *sctx)
703 {
704 if (!sctx->need_check_render_feedback)
705 return;
706
707 /* There is no render feedback if color writes are disabled.
708 * (e.g. a pixel shader with image stores)
709 */
710 if (!si_get_total_colormask(sctx))
711 return;
712
713 for (int i = 0; i < SI_NUM_GRAPHICS_SHADERS; ++i) {
714 if (!sctx->shaders[i].cso)
715 continue;
716
717 struct si_shader_info *info = &sctx->shaders[i].cso->info;
718 si_check_render_feedback_images(sctx, &sctx->images[i],
719 u_bit_consecutive(0, info->base.num_images));
720 si_check_render_feedback_textures(sctx, &sctx->samplers[i],
721 info->base.textures_used[0]);
722 }
723
724 si_check_render_feedback_resident_images(sctx);
725 si_check_render_feedback_resident_textures(sctx);
726
727 sctx->need_check_render_feedback = false;
728 }
729
si_decompress_resident_textures(struct si_context * sctx)730 static void si_decompress_resident_textures(struct si_context *sctx)
731 {
732 util_dynarray_foreach (&sctx->resident_tex_needs_color_decompress, struct si_texture_handle *,
733 tex_handle) {
734 struct pipe_sampler_view *view = (*tex_handle)->view;
735 struct si_texture *tex = (struct si_texture *)view->texture;
736
737 si_decompress_color_texture(sctx, tex, view->u.tex.first_level, view->u.tex.last_level,
738 false);
739 }
740
741 util_dynarray_foreach (&sctx->resident_tex_needs_depth_decompress, struct si_texture_handle *,
742 tex_handle) {
743 struct pipe_sampler_view *view = (*tex_handle)->view;
744 struct si_sampler_view *sview = (struct si_sampler_view *)view;
745 struct si_texture *tex = (struct si_texture *)view->texture;
746
747 si_decompress_depth(sctx, tex, sview->is_stencil_sampler ? PIPE_MASK_S : PIPE_MASK_Z,
748 view->u.tex.first_level, view->u.tex.last_level, 0,
749 util_max_layer(&tex->buffer.b.b, view->u.tex.first_level));
750 }
751 }
752
si_decompress_resident_images(struct si_context * sctx)753 static void si_decompress_resident_images(struct si_context *sctx)
754 {
755 util_dynarray_foreach (&sctx->resident_img_needs_color_decompress, struct si_image_handle *,
756 img_handle) {
757 struct pipe_image_view *view = &(*img_handle)->view;
758 struct si_texture *tex = (struct si_texture *)view->resource;
759
760 si_decompress_color_texture(sctx, tex, view->u.tex.level, view->u.tex.level,
761 view->access & PIPE_IMAGE_ACCESS_WRITE);
762 }
763 }
764
si_decompress_textures(struct si_context * sctx,unsigned shader_mask)765 void si_decompress_textures(struct si_context *sctx, unsigned shader_mask)
766 {
767 unsigned compressed_colortex_counter, mask;
768 bool need_flush = false;
769
770 if (sctx->blitter_running)
771 return;
772
773 /* Update the compressed_colortex_mask if necessary. */
774 compressed_colortex_counter = p_atomic_read(&sctx->screen->compressed_colortex_counter);
775 if (compressed_colortex_counter != sctx->last_compressed_colortex_counter) {
776 sctx->last_compressed_colortex_counter = compressed_colortex_counter;
777 si_update_needs_color_decompress_masks(sctx);
778 }
779
780 /* Decompress color & depth textures if needed. */
781 mask = sctx->shader_needs_decompress_mask & shader_mask;
782 while (mask) {
783 unsigned i = u_bit_scan(&mask);
784
785 if (sctx->samplers[i].needs_depth_decompress_mask) {
786 need_flush |= si_decompress_sampler_depth_textures(sctx, &sctx->samplers[i]);
787 }
788 if (sctx->samplers[i].needs_color_decompress_mask) {
789 si_decompress_sampler_color_textures(sctx, &sctx->samplers[i]);
790 }
791 if (sctx->images[i].needs_color_decompress_mask) {
792 si_decompress_image_color_textures(sctx, &sctx->images[i]);
793 }
794 }
795
796 if (sctx->chip_class == GFX10_3 && need_flush) {
797 /* This fixes a corruption with the following sequence:
798 * - fast clear depth
799 * - decompress depth
800 * - draw
801 * (see https://gitlab.freedesktop.org/drm/amd/-/issues/1810#note_1170171)
802 */
803 sctx->b.flush(&sctx->b, NULL, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW);
804 }
805
806 if (shader_mask & u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS)) {
807 if (sctx->uses_bindless_samplers)
808 si_decompress_resident_textures(sctx);
809 if (sctx->uses_bindless_images)
810 si_decompress_resident_images(sctx);
811
812 if (sctx->ps_uses_fbfetch) {
813 struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
814 si_decompress_color_texture(sctx, (struct si_texture *)cb0->texture,
815 cb0->u.tex.first_layer, cb0->u.tex.last_layer, false);
816 }
817
818 si_check_render_feedback(sctx);
819 } else if (shader_mask & (1 << PIPE_SHADER_COMPUTE)) {
820 if (sctx->cs_shader_state.program->sel.info.uses_bindless_samplers)
821 si_decompress_resident_textures(sctx);
822 if (sctx->cs_shader_state.program->sel.info.uses_bindless_images)
823 si_decompress_resident_images(sctx);
824 }
825 }
826
827 /* Helper for decompressing a portion of a color or depth resource before
828 * blitting if any decompression is needed.
829 * The driver doesn't decompress resources automatically while u_blitter is
830 * rendering. */
si_decompress_subresource(struct pipe_context * ctx,struct pipe_resource * tex,unsigned planes,unsigned level,unsigned first_layer,unsigned last_layer)831 void si_decompress_subresource(struct pipe_context *ctx, struct pipe_resource *tex, unsigned planes,
832 unsigned level, unsigned first_layer, unsigned last_layer)
833 {
834 struct si_context *sctx = (struct si_context *)ctx;
835 struct si_texture *stex = (struct si_texture *)tex;
836
837 if (stex->db_compatible) {
838 planes &= PIPE_MASK_Z | PIPE_MASK_S;
839
840 if (!stex->surface.has_stencil)
841 planes &= ~PIPE_MASK_S;
842
843 /* If we've rendered into the framebuffer and it's a blitting
844 * source, make sure the decompression pass is invoked
845 * by dirtying the framebuffer.
846 */
847 if (sctx->framebuffer.state.zsbuf && sctx->framebuffer.state.zsbuf->u.tex.level == level &&
848 sctx->framebuffer.state.zsbuf->texture == tex)
849 si_update_fb_dirtiness_after_rendering(sctx);
850
851 si_decompress_depth(sctx, stex, planes, level, level, first_layer, last_layer);
852 } else if (stex->surface.fmask_size || stex->cmask_buffer ||
853 vi_dcc_enabled(stex, level)) {
854 /* If we've rendered into the framebuffer and it's a blitting
855 * source, make sure the decompression pass is invoked
856 * by dirtying the framebuffer.
857 */
858 for (unsigned i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
859 if (sctx->framebuffer.state.cbufs[i] &&
860 sctx->framebuffer.state.cbufs[i]->u.tex.level == level &&
861 sctx->framebuffer.state.cbufs[i]->texture == tex) {
862 si_update_fb_dirtiness_after_rendering(sctx);
863 break;
864 }
865 }
866
867 si_blit_decompress_color(sctx, stex, level, level, first_layer, last_layer, false, false);
868 }
869 }
870
871 struct texture_orig_info {
872 unsigned format;
873 unsigned width0;
874 unsigned height0;
875 unsigned npix_x;
876 unsigned npix_y;
877 unsigned npix0_x;
878 unsigned npix0_y;
879 };
880
si_use_compute_copy_for_float_formats(struct si_context * sctx,struct pipe_resource * texture,unsigned level)881 static void si_use_compute_copy_for_float_formats(struct si_context *sctx,
882 struct pipe_resource *texture,
883 unsigned level) {
884 struct si_texture *tex = (struct si_texture *)texture;
885
886 /* If we are uploading into FP16 or R11G11B10_FLOAT via a blit, CB clobbers NaNs,
887 * so in order to preserve them exactly, we have to use the compute blit.
888 * The compute blit is used only when the destination doesn't have DCC, so
889 * disable it here, which is kinda a hack.
890 * If we are uploading into 32-bit floats with DCC via a blit, NaNs will also get
891 * lost so we need to disable DCC as well.
892 *
893 * This makes KHR-GL45.texture_view.view_classes pass on gfx9.
894 */
895 if (vi_dcc_enabled(tex, level) &&
896 util_format_is_float(texture->format) &&
897 sctx->chip_class < GFX10) {
898 si_texture_disable_dcc(sctx, tex);
899 }
900 }
901
si_resource_copy_region(struct pipe_context * ctx,struct pipe_resource * dst,unsigned dst_level,unsigned dstx,unsigned dsty,unsigned dstz,struct pipe_resource * src,unsigned src_level,const struct pipe_box * src_box)902 void si_resource_copy_region(struct pipe_context *ctx, struct pipe_resource *dst,
903 unsigned dst_level, unsigned dstx, unsigned dsty, unsigned dstz,
904 struct pipe_resource *src, unsigned src_level,
905 const struct pipe_box *src_box)
906 {
907 struct si_context *sctx = (struct si_context *)ctx;
908 struct si_texture *ssrc = (struct si_texture *)src;
909 struct si_texture *sdst = (struct si_texture *)dst;
910 struct pipe_surface *dst_view, dst_templ;
911 struct pipe_sampler_view src_templ, *src_view;
912 unsigned dst_width, dst_height, src_width0, src_height0;
913 unsigned dst_width0, dst_height0, src_force_level = 0;
914 struct pipe_box sbox, dstbox;
915
916 /* Handle buffers first. */
917 if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
918 si_copy_buffer(sctx, dst, src, dstx, src_box->x, src_box->width, SI_OP_SYNC_BEFORE_AFTER);
919 return;
920 }
921
922 si_use_compute_copy_for_float_formats(sctx, dst, dst_level);
923
924 if (!util_format_is_compressed(src->format) && !util_format_is_compressed(dst->format) &&
925 !util_format_is_depth_or_stencil(src->format) && src->nr_samples <= 1 &&
926 /* DCC compression from image store is enabled for GFX10+. */
927 (!vi_dcc_enabled(sdst, dst_level) || sctx->chip_class >= GFX10) &&
928 !(dst->target != src->target &&
929 (src->target == PIPE_TEXTURE_1D_ARRAY || dst->target == PIPE_TEXTURE_1D_ARRAY))) {
930 si_compute_copy_image(sctx, dst, dst_level, src, src_level, dstx, dsty, dstz,
931 src_box, false, SI_OP_SYNC_BEFORE_AFTER);
932 return;
933 }
934
935 assert(u_max_sample(dst) == u_max_sample(src));
936
937 /* The driver doesn't decompress resources automatically while
938 * u_blitter is rendering. */
939 si_decompress_subresource(ctx, src, PIPE_MASK_RGBAZS, src_level, src_box->z,
940 src_box->z + src_box->depth - 1);
941
942 dst_width = u_minify(dst->width0, dst_level);
943 dst_height = u_minify(dst->height0, dst_level);
944 dst_width0 = dst->width0;
945 dst_height0 = dst->height0;
946 src_width0 = src->width0;
947 src_height0 = src->height0;
948
949 util_blitter_default_dst_texture(&dst_templ, dst, dst_level, dstz);
950 util_blitter_default_src_texture(sctx->blitter, &src_templ, src, src_level);
951
952 if (util_format_is_compressed(src->format) || util_format_is_compressed(dst->format)) {
953 unsigned blocksize = ssrc->surface.bpe;
954
955 if (blocksize == 8)
956 src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; /* 64-bit block */
957 else
958 src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; /* 128-bit block */
959 dst_templ.format = src_templ.format;
960
961 dst_width = util_format_get_nblocksx(dst->format, dst_width);
962 dst_height = util_format_get_nblocksy(dst->format, dst_height);
963 dst_width0 = util_format_get_nblocksx(dst->format, dst_width0);
964 dst_height0 = util_format_get_nblocksy(dst->format, dst_height0);
965 src_width0 = util_format_get_nblocksx(src->format, src_width0);
966 src_height0 = util_format_get_nblocksy(src->format, src_height0);
967
968 dstx = util_format_get_nblocksx(dst->format, dstx);
969 dsty = util_format_get_nblocksy(dst->format, dsty);
970
971 sbox.x = util_format_get_nblocksx(src->format, src_box->x);
972 sbox.y = util_format_get_nblocksy(src->format, src_box->y);
973 sbox.z = src_box->z;
974 sbox.width = util_format_get_nblocksx(src->format, src_box->width);
975 sbox.height = util_format_get_nblocksy(src->format, src_box->height);
976 sbox.depth = src_box->depth;
977 src_box = &sbox;
978
979 src_force_level = src_level;
980 } else if (!util_blitter_is_copy_supported(sctx->blitter, dst, src)) {
981 if (util_format_is_subsampled_422(src->format)) {
982 src_templ.format = PIPE_FORMAT_R8G8B8A8_UINT;
983 dst_templ.format = PIPE_FORMAT_R8G8B8A8_UINT;
984
985 dst_width = util_format_get_nblocksx(dst->format, dst_width);
986 dst_width0 = util_format_get_nblocksx(dst->format, dst_width0);
987 src_width0 = util_format_get_nblocksx(src->format, src_width0);
988
989 dstx = util_format_get_nblocksx(dst->format, dstx);
990
991 sbox = *src_box;
992 sbox.x = util_format_get_nblocksx(src->format, src_box->x);
993 sbox.width = util_format_get_nblocksx(src->format, src_box->width);
994 src_box = &sbox;
995 } else {
996 unsigned blocksize = ssrc->surface.bpe;
997
998 switch (blocksize) {
999 case 1:
1000 dst_templ.format = PIPE_FORMAT_R8_UNORM;
1001 src_templ.format = PIPE_FORMAT_R8_UNORM;
1002 break;
1003 case 2:
1004 dst_templ.format = PIPE_FORMAT_R8G8_UNORM;
1005 src_templ.format = PIPE_FORMAT_R8G8_UNORM;
1006 break;
1007 case 4:
1008 dst_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM;
1009 src_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM;
1010 break;
1011 case 8:
1012 dst_templ.format = PIPE_FORMAT_R16G16B16A16_UINT;
1013 src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT;
1014 break;
1015 case 16:
1016 dst_templ.format = PIPE_FORMAT_R32G32B32A32_UINT;
1017 src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT;
1018 break;
1019 default:
1020 fprintf(stderr, "Unhandled format %s with blocksize %u\n",
1021 util_format_short_name(src->format), blocksize);
1022 assert(0);
1023 }
1024 }
1025 }
1026
1027 /* SNORM8 blitting has precision issues on some chips. Use the SINT
1028 * equivalent instead, which doesn't force DCC decompression.
1029 */
1030 if (util_format_is_snorm8(dst_templ.format)) {
1031 dst_templ.format = src_templ.format = util_format_snorm8_to_sint8(dst_templ.format);
1032 }
1033
1034 vi_disable_dcc_if_incompatible_format(sctx, dst, dst_level, dst_templ.format);
1035 vi_disable_dcc_if_incompatible_format(sctx, src, src_level, src_templ.format);
1036
1037 /* Initialize the surface. */
1038 dst_view = si_create_surface_custom(ctx, dst, &dst_templ, dst_width0, dst_height0, dst_width,
1039 dst_height);
1040
1041 /* Initialize the sampler view. */
1042 src_view =
1043 si_create_sampler_view_custom(ctx, src, &src_templ, src_width0, src_height0, src_force_level);
1044
1045 u_box_3d(dstx, dsty, dstz, abs(src_box->width), abs(src_box->height), abs(src_box->depth),
1046 &dstbox);
1047
1048 /* Copy. */
1049 si_blitter_begin(sctx, SI_COPY);
1050 util_blitter_blit_generic(sctx->blitter, dst_view, &dstbox, src_view, src_box, src_width0,
1051 src_height0, PIPE_MASK_RGBAZS, PIPE_TEX_FILTER_NEAREST, NULL, false, false);
1052 si_blitter_end(sctx);
1053
1054 pipe_surface_reference(&dst_view, NULL);
1055 pipe_sampler_view_reference(&src_view, NULL);
1056 }
1057
si_do_CB_resolve(struct si_context * sctx,const struct pipe_blit_info * info,struct pipe_resource * dst,unsigned dst_level,unsigned dst_z,enum pipe_format format)1058 static void si_do_CB_resolve(struct si_context *sctx, const struct pipe_blit_info *info,
1059 struct pipe_resource *dst, unsigned dst_level, unsigned dst_z,
1060 enum pipe_format format)
1061 {
1062 /* Required before and after CB_RESOLVE. */
1063 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB;
1064
1065 si_blitter_begin(
1066 sctx, SI_COLOR_RESOLVE | (info->render_condition_enable ? 0 : SI_DISABLE_RENDER_COND));
1067 util_blitter_custom_resolve_color(sctx->blitter, dst, dst_level, dst_z, info->src.resource,
1068 info->src.box.z, ~0, sctx->custom_blend_resolve, format);
1069 si_blitter_end(sctx);
1070
1071 /* Flush caches for possible texturing. */
1072 si_make_CB_shader_coherent(sctx, 1, false, true /* no DCC */);
1073 }
1074
resolve_formats_compatible(enum pipe_format src,enum pipe_format dst,bool src_swaps_rgb_to_bgr,bool * need_rgb_to_bgr)1075 static bool resolve_formats_compatible(enum pipe_format src, enum pipe_format dst,
1076 bool src_swaps_rgb_to_bgr, bool *need_rgb_to_bgr)
1077 {
1078 *need_rgb_to_bgr = false;
1079
1080 if (src_swaps_rgb_to_bgr) {
1081 /* We must only check the swapped format. */
1082 enum pipe_format swapped_src = util_format_rgb_to_bgr(src);
1083 assert(swapped_src);
1084 return util_is_format_compatible(util_format_description(swapped_src),
1085 util_format_description(dst));
1086 }
1087
1088 if (util_is_format_compatible(util_format_description(src), util_format_description(dst)))
1089 return true;
1090
1091 enum pipe_format swapped_src = util_format_rgb_to_bgr(src);
1092 *need_rgb_to_bgr = util_is_format_compatible(util_format_description(swapped_src),
1093 util_format_description(dst));
1094 return *need_rgb_to_bgr;
1095 }
1096
do_hardware_msaa_resolve(struct pipe_context * ctx,const struct pipe_blit_info * info)1097 static bool do_hardware_msaa_resolve(struct pipe_context *ctx, const struct pipe_blit_info *info)
1098 {
1099 struct si_context *sctx = (struct si_context *)ctx;
1100 struct si_texture *src = (struct si_texture *)info->src.resource;
1101 struct si_texture *dst = (struct si_texture *)info->dst.resource;
1102 ASSERTED struct si_texture *stmp;
1103 unsigned dst_width = u_minify(info->dst.resource->width0, info->dst.level);
1104 unsigned dst_height = u_minify(info->dst.resource->height0, info->dst.level);
1105 enum pipe_format format = info->src.format;
1106 struct pipe_resource *tmp, templ;
1107 struct pipe_blit_info blit;
1108
1109 /* Check basic requirements for hw resolve. */
1110 if (!(info->src.resource->nr_samples > 1 && info->dst.resource->nr_samples <= 1 &&
1111 !util_format_is_pure_integer(format) && !util_format_is_depth_or_stencil(format) &&
1112 util_max_layer(info->src.resource, 0) == 0))
1113 return false;
1114
1115 /* Hardware MSAA resolve doesn't work if SPI format = NORM16_ABGR and
1116 * the format is R16G16. Use R16A16, which does work.
1117 */
1118 if (format == PIPE_FORMAT_R16G16_UNORM)
1119 format = PIPE_FORMAT_R16A16_UNORM;
1120 if (format == PIPE_FORMAT_R16G16_SNORM)
1121 format = PIPE_FORMAT_R16A16_SNORM;
1122
1123 bool need_rgb_to_bgr = false;
1124
1125 /* Check the remaining requirements for hw resolve. */
1126 if (util_max_layer(info->dst.resource, info->dst.level) == 0 && !info->scissor_enable &&
1127 (info->mask & PIPE_MASK_RGBA) == PIPE_MASK_RGBA &&
1128 resolve_formats_compatible(info->src.format, info->dst.format,
1129 src->swap_rgb_to_bgr, &need_rgb_to_bgr) &&
1130 dst_width == info->src.resource->width0 && dst_height == info->src.resource->height0 &&
1131 info->dst.box.x == 0 && info->dst.box.y == 0 && info->dst.box.width == dst_width &&
1132 info->dst.box.height == dst_height && info->dst.box.depth == 1 && info->src.box.x == 0 &&
1133 info->src.box.y == 0 && info->src.box.width == dst_width &&
1134 info->src.box.height == dst_height && info->src.box.depth == 1 && !dst->surface.is_linear &&
1135 (!dst->cmask_buffer || !dst->dirty_level_mask)) { /* dst cannot be fast-cleared */
1136 /* Check the remaining constraints. */
1137 if (src->surface.micro_tile_mode != dst->surface.micro_tile_mode ||
1138 need_rgb_to_bgr) {
1139 /* The next fast clear will switch to this mode to
1140 * get direct hw resolve next time if the mode is
1141 * different now.
1142 *
1143 * TODO-GFX10: This does not work in GFX10 because MSAA
1144 * is restricted to 64KB_R_X and 64KB_Z_X swizzle modes.
1145 * In some cases we could change the swizzle of the
1146 * destination texture instead, but the more general
1147 * solution is to implement compute shader resolve.
1148 */
1149 if (src->surface.micro_tile_mode != dst->surface.micro_tile_mode)
1150 src->last_msaa_resolve_target_micro_mode = dst->surface.micro_tile_mode;
1151 if (need_rgb_to_bgr)
1152 src->swap_rgb_to_bgr_on_next_clear = true;
1153
1154 goto resolve_to_temp;
1155 }
1156
1157 /* Resolving into a surface with DCC is unsupported. Since
1158 * it's being overwritten anyway, clear it to uncompressed.
1159 * This is still the fastest codepath even with this clear.
1160 */
1161 if (vi_dcc_enabled(dst, info->dst.level)) {
1162 struct si_clear_info clear_info;
1163
1164 if (!vi_dcc_get_clear_info(sctx, dst, info->dst.level, DCC_UNCOMPRESSED, &clear_info))
1165 goto resolve_to_temp;
1166
1167 si_execute_clears(sctx, &clear_info, 1, SI_CLEAR_TYPE_DCC);
1168 dst->dirty_level_mask &= ~(1 << info->dst.level);
1169 }
1170
1171 /* Resolve directly from src to dst. */
1172 si_do_CB_resolve(sctx, info, info->dst.resource, info->dst.level, info->dst.box.z, format);
1173 return true;
1174 }
1175
1176 resolve_to_temp:
1177 /* Shader-based resolve is VERY SLOW. Instead, resolve into
1178 * a temporary texture and blit.
1179 */
1180 memset(&templ, 0, sizeof(templ));
1181 templ.target = PIPE_TEXTURE_2D;
1182 templ.format = info->src.resource->format;
1183 templ.width0 = info->src.resource->width0;
1184 templ.height0 = info->src.resource->height0;
1185 templ.depth0 = 1;
1186 templ.array_size = 1;
1187 templ.usage = PIPE_USAGE_DEFAULT;
1188 templ.flags = SI_RESOURCE_FLAG_FORCE_MSAA_TILING | SI_RESOURCE_FLAG_FORCE_MICRO_TILE_MODE |
1189 SI_RESOURCE_FLAG_MICRO_TILE_MODE_SET(src->surface.micro_tile_mode) |
1190 SI_RESOURCE_FLAG_DISABLE_DCC | SI_RESOURCE_FLAG_DRIVER_INTERNAL;
1191
1192 /* The src and dst microtile modes must be the same. */
1193 if (sctx->chip_class <= GFX8 && src->surface.micro_tile_mode == RADEON_MICRO_MODE_DISPLAY)
1194 templ.bind = PIPE_BIND_SCANOUT;
1195 else
1196 templ.bind = 0;
1197
1198 tmp = ctx->screen->resource_create(ctx->screen, &templ);
1199 if (!tmp)
1200 return false;
1201 stmp = (struct si_texture *)tmp;
1202 /* Match the channel order of src. */
1203 stmp->swap_rgb_to_bgr = src->swap_rgb_to_bgr;
1204
1205 assert(!stmp->surface.is_linear);
1206 assert(src->surface.micro_tile_mode == stmp->surface.micro_tile_mode);
1207
1208 /* resolve */
1209 si_do_CB_resolve(sctx, info, tmp, 0, 0, format);
1210
1211 /* blit */
1212 blit = *info;
1213 blit.src.resource = tmp;
1214 blit.src.box.z = 0;
1215
1216 si_blitter_begin(sctx, SI_BLIT | (info->render_condition_enable ? 0 : SI_DISABLE_RENDER_COND));
1217 util_blitter_blit(sctx->blitter, &blit);
1218 si_blitter_end(sctx);
1219
1220 pipe_resource_reference(&tmp, NULL);
1221 return true;
1222 }
1223
si_blit(struct pipe_context * ctx,const struct pipe_blit_info * info)1224 static void si_blit(struct pipe_context *ctx, const struct pipe_blit_info *info)
1225 {
1226 struct si_context *sctx = (struct si_context *)ctx;
1227 struct si_texture *sdst = (struct si_texture *)info->dst.resource;
1228
1229 if (do_hardware_msaa_resolve(ctx, info)) {
1230 return;
1231 }
1232
1233 if (info->is_dri_blit_image && sdst->surface.is_linear &&
1234 sctx->chip_class >= GFX7 && sdst->surface.flags & RADEON_SURF_IMPORTED) {
1235 struct si_texture *ssrc = (struct si_texture *)info->src.resource;
1236 /* Use SDMA or async compute when copying to a DRI_PRIME imported linear surface. */
1237 bool async_copy = info->dst.box.x == 0 && info->dst.box.y == 0 && info->dst.box.z == 0 &&
1238 info->src.box.x == 0 && info->src.box.y == 0 && info->src.box.z == 0 &&
1239 info->dst.level == 0 && info->src.level == 0 &&
1240 info->src.box.width == info->dst.resource->width0 &&
1241 info->src.box.height == info->dst.resource->height0 &&
1242 info->src.box.depth == 1 && util_can_blit_via_copy_region(info, true);
1243 /* Try SDMA first... */
1244 /* TODO: figure out why SDMA copies are slow on GFX10_3 */
1245 if (async_copy && sctx->chip_class < GFX10_3 && si_sdma_copy_image(sctx, sdst, ssrc))
1246 return;
1247
1248 /* ... and use async compute as the fallback. */
1249 if (async_copy) {
1250 struct si_screen *sscreen = sctx->screen;
1251
1252 simple_mtx_lock(&sscreen->async_compute_context_lock);
1253 if (!sscreen->async_compute_context)
1254 si_init_aux_async_compute_ctx(sscreen);
1255
1256 if (sscreen->async_compute_context) {
1257 si_compute_copy_image((struct si_context*)sctx->screen->async_compute_context,
1258 info->dst.resource, 0, info->src.resource, 0, 0, 0, 0,
1259 &info->src.box, false, 0);
1260 si_flush_gfx_cs((struct si_context*)sctx->screen->async_compute_context, 0, NULL);
1261 simple_mtx_unlock(&sscreen->async_compute_context_lock);
1262 return;
1263 }
1264
1265 simple_mtx_unlock(&sscreen->async_compute_context_lock);
1266 }
1267 }
1268
1269 if (unlikely(sctx->thread_trace_enabled))
1270 sctx->sqtt_next_event = EventCmdCopyImage;
1271
1272 /* Using compute for copying to a linear texture in GTT is much faster than
1273 * going through RBs (render backends). This improves DRI PRIME performance.
1274 */
1275 if (util_can_blit_via_copy_region(info, false)) {
1276 si_resource_copy_region(ctx, info->dst.resource, info->dst.level,
1277 info->dst.box.x, info->dst.box.y, info->dst.box.z,
1278 info->src.resource, info->src.level, &info->src.box);
1279 return;
1280 }
1281
1282 assert(util_blitter_is_blit_supported(sctx->blitter, info));
1283
1284 /* The driver doesn't decompress resources automatically while
1285 * u_blitter is rendering. */
1286 vi_disable_dcc_if_incompatible_format(sctx, info->src.resource, info->src.level,
1287 info->src.format);
1288 vi_disable_dcc_if_incompatible_format(sctx, info->dst.resource, info->dst.level,
1289 info->dst.format);
1290 si_decompress_subresource(ctx, info->src.resource, PIPE_MASK_RGBAZS, info->src.level,
1291 info->src.box.z, info->src.box.z + info->src.box.depth - 1);
1292
1293 if (unlikely(sctx->thread_trace_enabled))
1294 sctx->sqtt_next_event = EventCmdBlitImage;
1295
1296 si_blitter_begin(sctx, SI_BLIT | (info->render_condition_enable ? 0 : SI_DISABLE_RENDER_COND));
1297 util_blitter_blit(sctx->blitter, info);
1298 si_blitter_end(sctx);
1299 }
1300
si_generate_mipmap(struct pipe_context * ctx,struct pipe_resource * tex,enum pipe_format format,unsigned base_level,unsigned last_level,unsigned first_layer,unsigned last_layer)1301 static bool si_generate_mipmap(struct pipe_context *ctx, struct pipe_resource *tex,
1302 enum pipe_format format, unsigned base_level, unsigned last_level,
1303 unsigned first_layer, unsigned last_layer)
1304 {
1305 struct si_context *sctx = (struct si_context *)ctx;
1306 struct si_texture *stex = (struct si_texture *)tex;
1307
1308 if (!util_blitter_is_copy_supported(sctx->blitter, tex, tex))
1309 return false;
1310
1311 /* The driver doesn't decompress resources automatically while
1312 * u_blitter is rendering. */
1313 vi_disable_dcc_if_incompatible_format(sctx, tex, base_level, format);
1314 si_decompress_subresource(ctx, tex, PIPE_MASK_RGBAZS, base_level, first_layer, last_layer);
1315
1316 /* Clear dirty_level_mask for the levels that will be overwritten. */
1317 assert(base_level < last_level);
1318 stex->dirty_level_mask &= ~u_bit_consecutive(base_level + 1, last_level - base_level);
1319
1320 sctx->generate_mipmap_for_depth = stex->is_depth;
1321
1322 si_blitter_begin(sctx, SI_BLIT | SI_DISABLE_RENDER_COND);
1323 util_blitter_generate_mipmap(sctx->blitter, tex, format, base_level, last_level, first_layer,
1324 last_layer);
1325 si_blitter_end(sctx);
1326
1327 sctx->generate_mipmap_for_depth = false;
1328 return true;
1329 }
1330
si_flush_resource(struct pipe_context * ctx,struct pipe_resource * res)1331 static void si_flush_resource(struct pipe_context *ctx, struct pipe_resource *res)
1332 {
1333 struct si_context *sctx = (struct si_context *)ctx;
1334 struct si_texture *tex = (struct si_texture *)res;
1335
1336 assert(res->target != PIPE_BUFFER);
1337
1338 if (!tex->is_depth && (tex->cmask_buffer || vi_dcc_enabled(tex, 0))) {
1339 si_blit_decompress_color(sctx, tex, 0, res->last_level, 0, util_max_layer(res, 0),
1340 false, false);
1341
1342 if (tex->surface.display_dcc_offset && tex->displayable_dcc_dirty) {
1343 si_retile_dcc(sctx, tex);
1344 tex->displayable_dcc_dirty = false;
1345 }
1346 }
1347 }
1348
si_flush_implicit_resources(struct si_context * sctx)1349 void si_flush_implicit_resources(struct si_context *sctx)
1350 {
1351 hash_table_foreach(sctx->dirty_implicit_resources, entry) {
1352 si_flush_resource(&sctx->b, entry->data);
1353 pipe_resource_reference((struct pipe_resource **)&entry->data, NULL);
1354 }
1355 _mesa_hash_table_clear(sctx->dirty_implicit_resources, NULL);
1356 }
1357
si_decompress_dcc(struct si_context * sctx,struct si_texture * tex)1358 void si_decompress_dcc(struct si_context *sctx, struct si_texture *tex)
1359 {
1360 assert(!tex->is_depth);
1361
1362 /* If graphics is disabled, we can't decompress DCC, but it shouldn't
1363 * be compressed either. The caller should simply discard it.
1364 */
1365 if (!tex->surface.meta_offset || !sctx->has_graphics)
1366 return;
1367
1368 if (sctx->chip_class == GFX8 || tex->buffer.b.b.nr_storage_samples >= 2) {
1369 si_blit_decompress_color(sctx, tex, 0, tex->buffer.b.b.last_level, 0,
1370 util_max_layer(&tex->buffer.b.b, 0), true, false);
1371 } else {
1372 struct pipe_resource *ptex = &tex->buffer.b.b;
1373 assert(ptex->nr_storage_samples <= 1);
1374
1375 /* DCC decompression using a compute shader. */
1376 for (unsigned level = 0; level < tex->surface.num_meta_levels; level++) {
1377 struct pipe_box box;
1378
1379 u_box_3d(0, 0, 0, u_minify(ptex->width0, level),
1380 u_minify(ptex->height0, level),
1381 util_num_layers(ptex, level), &box);
1382 si_compute_copy_image(sctx, ptex, level, ptex, level, 0, 0, 0, &box, true,
1383 /* Sync before the first copy and after the last copy */
1384 (level == 0 ? SI_OP_SYNC_BEFORE : 0) |
1385 (level == tex->surface.num_meta_levels - 1 ? SI_OP_SYNC_AFTER : 0));
1386 }
1387
1388 /* Now clear DCC metadata to uncompressed.
1389 *
1390 * This uses SI_COMPUTE_CLEAR_METHOD to avoid a failure when running this
1391 * deqp caselist on gfx10:
1392 * dEQP-GLES31.functional.image_load_store.2d.format_reinterpret.rgba32f_rgba32ui
1393 * dEQP-GLES31.functional.image_load_store.2d.format_reinterpret.rgba32f_rgba32i
1394 */
1395 uint32_t clear_value = DCC_UNCOMPRESSED;
1396 si_clear_buffer(sctx, ptex, tex->surface.meta_offset,
1397 tex->surface.meta_size, &clear_value, 4, SI_OP_SYNC_AFTER,
1398 SI_COHERENCY_CB_META, SI_COMPUTE_CLEAR_METHOD);
1399 si_mark_display_dcc_dirty(sctx, tex);
1400
1401 /* Clearing DCC metadata requires flushing L2 and invalidating L2 metadata to make
1402 * the metadata visible to L2 caches. This is because clear_buffer uses plain stores
1403 * that can go to different L2 channels than where L2 metadata caches expect them.
1404 * This is not done for fast clears because plain stores are visible to CB/DB. Only
1405 * L2 metadata caches have the problem.
1406 */
1407 sctx->flags |= SI_CONTEXT_WB_L2 | SI_CONTEXT_INV_L2_METADATA;
1408 }
1409 }
1410
si_init_blit_functions(struct si_context * sctx)1411 void si_init_blit_functions(struct si_context *sctx)
1412 {
1413 sctx->b.resource_copy_region = si_resource_copy_region;
1414
1415 if (sctx->has_graphics) {
1416 sctx->b.blit = si_blit;
1417 sctx->b.flush_resource = si_flush_resource;
1418 sctx->b.generate_mipmap = si_generate_mipmap;
1419 }
1420 }
1421